Files

Garry Tan ede85e2e8c v0.48.3.0 fix: consistent read policies and reliable verification (#4941 )

* fix(ci): require executed verification and isolate routing fixtures

Remove cached-success bypasses and preserve failed, cancelled, or unexpectedly
skipped execution lanes. Require scheduled full-corpus E2E and actual PgBouncer
execution, retain logs, and install locked dependencies consistently.

Isolate routing fixture provider state and home directories, disable Bun env-file
loading in children, and report bounded redacted failures. Keep coverage
percentages advisory under the existing graduation rules.

* fix(mcp): limit authentication replay and propagate cancellation

Refresh and reconstruct the client once only for the SDK's typed HTTP 401.
Preserve operation errors and treat other transport failures as terminal.
Carry cancellation through discovery, token minting, connection and tool calls,
and close clients after failed initialization or aborted requests.

* fix(privacy): enforce scoped reads and certify safe search indexes

Authorize concrete rows and contributing pages throughout both engines,
history, graphs, search enrichment, and holder-dependent calculations.
Normalize stored prose before strict protected-section filtering and replace
chunks atomically before admitting them to remote retrieval.

Disable shared semantic result caching and restrict stored contradiction
reports to trusted unscoped local callers. Keep code inspection local-only
until its graph expansion can enforce the same complete read policy.

Cover source collisions, history, holder restrictions, index rebuilds,
normalization, containment, and real HTTP/stdio and database parity journeys.

* v0.48.3.0 fix: consistent read policies and reliable verification

Synchronize release metadata and generated distribution stamps. Document the
required index rebuild, temporary retrieval availability, provider costs,
remote salience behavior, and deferred provenance and code-read follow-ups.

* v0.48.3.0 docs: align retrieval and upgrade guidance

* fix(test): isolate nested E2E coverage output

2026-09-06 12:00:51 -07:00

3.3 KiB

Raw Permalink Blame History

v0.48.3.0: verify access rules and temporary search behavior

This release applies access rules consistently and improves remote-call and verification behavior. It requires no database schema migration or new service.

Verify the installed version and brain health:

gbrain --version
gbrain doctor
gbrain search modes
gbrain cache stats

Explain the temporary semantic result-cache change to the operator. Searches use fresh retrieval even when caching is configured. Cache statistics report enabled: false; stored rows and maintenance commands remain available. Repeated searches may take longer and use more provider tokens. Do not change provider credentials or promise that a cache setting can restore caching.
Stored contradiction reports are temporarily available only to trusted local callers without a source filter. Remote and source-scoped requests receive an availability note. Do not broaden a caller's source permissions to work around that response.
Explain any changed remote salience scores using the caller's permitted active takes. Restricted scores omit stored emotional weight and use page updates for recent inclusion. Existing trusted local callers without holder restrictions retain their formulas.
Existing chunks are withheld from chunk-based remote retrieval until rebuilt with the corrected indexer. This applies to public pages too. Direct page reads remain available under the connection's access rules. Preview the existing rebuild:
```
gbrain reindex --markdown --dry-run --no-embed
```
For a rebuild without embedding-provider calls:
```
gbrain reindex --markdown --no-embed
```
This replaces processed pages' chunks and clears their previous vectors. Keyword retrieval becomes available after rebuilding; semantic retrieval needs fresh embeddings. Use the normal gbrain reindex --markdown path when embedding is authorized, or restore vectors later with gbrain embed --stale. Explain provider usage before running either embedding path. Preserve the existing post-upgrade cost confirmation. Do not manually change chunker version fields to suppress pending work.

Code pages use the existing gbrain reindex-code --force --no-embed path; add --dry-run to preview it. Image pages need their original files reimported so OCR and visual indexing are rebuilt together. Explain any OCR or embedding-provider usage before that import. Images with protected OCR sections remain unavailable to remote chunk retrieval; update the source image before reimporting it. Direct page access still follows normal policy.
Code-inspection operations (code_def, code_refs, code_callers, code_callees, code_blast, code_flow) are temporarily available only to trusted local callers. Remote callers receive an explicit availability error. Rebuilding chunks does not lift this separate restriction. Remote search omits optional code-graph expansion; ordinary authorized text search remains available after the index rebuild. Use the trusted local CLI when local access is available, without widening a remote token's permissions.

No new scheduled job or host configuration is required. Follow the exact remediation printed by gbrain doctor if it finds an independent setup issue.

3.3 KiB Raw Permalink Blame History

v0.48.3.0: verify access rules and temporary search behavior

3.3 KiB

Raw Permalink Blame History